Swan(sea) Song – personal research during my six years at Swansea ... and bey...
Style based
1. d.tour: Style-based
Exploration of
Design Example Galleries
Author: Daniel Ritchie, Ankita Arvind
Kejriwal, Scott R. Klemmer
Stanford University HCI Group
Department of Computer Science
Stanford, CA, USA
Source: UIST’11, October 1619,
2011, Santa Barbara, CA, USA. ACM
Advisor: Geeng Neng You
Speaker: Ching Jung Chang
http://www.youtube.com/watch?v=g4ggoccLOyI&featur
e=related
1
2. INTRODUCTION
• An interface and interaction techniques for style-
based gallery exploration.
• Introduce exploratory techniques for finding
relevant and inspiring design examples.
• A curated database of Web pages .
• Methods for extracting and analyzing design
features from a curated corpus of Web pages.
• A recommender system that processes queries
and returns relevant pages from the database.
• It decreases the gulfs of execution and evaluation
for design example-finding.
2
3. Web interface : (a) Search Results, (b) Search box,
(c) Current Query, (d) Ambient Style Display, (e) Bookmarks
3
4. USER INTERFACE
• (a) Search Results: Presents search
results as interactive thumbnails.
• query interface by providing ‘Show more
like’ Show less like’ clicking a page
thumbnail to expand the page to full size
bookmark and add notes to the page.
• (b) Search box: Supports query by text
string, color keywords, style terms, or any
combination of the three. (colorful image-
heavy)
4
5. Searching for ‘colorful image heavy’ retrieves
colorful pages that make heavy use of images. 5
6. • (c) Current Query: This right-hand
pane displays the keyword and page
queries that specify the current state
of the design gallery.
• Building simple interfaces for users to
convey rationale is an important open
issue in recommender systems.
• (d) Ambient Style Display: When a
user hovers over a page, the display
provides instant feedback by showing
stylistic attributes .
• (e) Bookmarks: allows the user to
maintain a collection of interesting
pages across several search sessions.
6
7. The Ambient Style Display shows bar chart visualizes the
aggregate style attributes of the current example queries.
When the user mouse over a page thumbnail, it instead
7
shows that page’s stylistic attributes.
8. IMPLEMENTATION
• Database
• The current prototype includes around
300 Web pages listed on sites such as
webbyawards.com, designmeltdown.co
m, ilovetypography.com, sixrevisions.co
m, and the Alexa Top 100.
• The complete contents of the database:
bricolage.stanford.edu/database.
8
9. • System Architecture:
• All utilities for processing pages and
extracting features are implemented in:
• C++ with the Qt framework.
• QtWebKit API to interface with the
WebKit browser engine.
• Search interface : HTML, JavaScript, CSS.
• Supports HTML5-compliant browsers :
Firefox, Safari, and Chrome.
• Web server: JavaScript using node.js
9
10. FEATURE EXTRACTION
• Focuse on global pages features that
describe a design’s use of
space, color, text, and images.
• These features compose to predict
higher-level stylistic attributes.
• Many of these features are derived from
the page’s Document Object Model
(DOM) tree.
10
11. http://www.w3schools.com/
htmldom/dom_nodetree.asp
Parents
Children Siblings
• The HTML DOM views an HTML document as a tree
structure. The tree structure is called a node-tree.
• The tree starts at the root node and branches out to the
text nodes at the lowest level of the tree.
• Node Parents, Children, and Siblings. The nodes in the
node tree have a hierarchical relationship to each other.
11
12. <html>
<head>
<title>Hello DOM </title>
</head>
<body>
<h1>DOM Lesson
one</h1>
<p>Hello world!</p>
</body>
</html>
The <html> node has no parent node; it is the root node.
The <html> node has two child nodes; <head> and <body>
The <head> node has one child node; the <title> node
The <title> node also has one child node; the text node “Hello DOM "
The <h1> and <p> nodes are siblings, and both child nodes of <body>
12
13. Principles from the design literature
Use of DOM tree depth (min, max, mean, stddev), # of
Space DOM leaf nodes, document width/height, amount
of separation between content blocks,
foreground/background ratio, overlapping element
area, connected components in rendered image (#,
minarea, maxarea, meanarea)
Use of Color (mean, stddev), saturation (min, max, mean,
Color stddev), value (min, max, mean, stddev), # of
colors (in DOM, in rendered page), most dominant
color, most dominant text color, # of dominant
colors, text-to-background contrast (min, max,
mean, stddev), histogram
Use of # of words in page, # of words per block (min,
Text max, mean, stddev), # of fonts, font size (min, max,
mode, mean)
Use of # of images, aspectRatio (min, max, mean,
Images stddev), area (min, max, mean, stddev), complexity
(min, max, mean, stddev) 13
14. • Other features require Segmentation
• Apply the Bento page segmentation algorithm
to ensure that the DOM tree accurately
reflects the page’s visual layout.
• Bento: data collection study, the mapping
algorithm, and the machine learning method.
• The system also computes a color histogram of
the rendered page, quantizing the image into
256 colors that evenly sample the RGB color
cube.
• Additional features for the local structure web
pages.
• Bricolage algorithm transfers the contents of a
source page into the style of a target page by
computing a mapping between the two pages’
DOM trees.
14
15. Left: The colored boxes illustrate Bento’s
segmentation.
Right: Bento’s output tree and associated
DOM nodes for this page.
15
16. Red
128- 192-
0-63 64-127
191 255
Original
0-63 43 78 18 0
64-127 45 67 33 2
Blue 128-
127 58 25 8
191
192-
140 47 47 13
256 colors • Color histogram in the
255 RGB color space
• A Red–Blue chromaticity histogram can be formed by
first normalizing color pixel values by dividing RGB
values by R+G+B.
• Quantizing the normalized R and B coordinates into N
bins each.
• A two-dimensional histogram of Red-Blue
chromaticity divide in to four bins (N=4) might yield a
histogram that looks like this table:
16
Reference: Wiki
18. QUERY PROCESSING
• Use the text contents of its pages and the
combined feature vectors described above to
respond to different types of queries.
• 1. Show More/Less Like’
• q weighted average of all feature vectors for pages
rated via ‘show more/less like.’
• M be the set of feature vectors for all ‘show more
like’ pages
• L be the set of vectors for all ‘show less like’ pages.
18
19. • 2. Text Keyword Search
• System computes a TF-IDF score for each word in the
page’s document text .
• It returns a ranked list of the pages whose combined
scores for the query terms are greater than zero.
• TF-IDF : Term Frequency–Inverse Document Frequency.
• Consider a document containing 100 words where in the
word colorful appears 3 times.
• Term frequency (TF) for colorful (3 / 100) = 0.03.
• Assume 10 million documents and colorful appears in
1,000 of these. DF=(10, 000 ,000 / 1, 000)
• Inverse document frequency is calculated as log(10, 000
,000 / 1, 000) =10,000= 4.
• tf–idf score : 0.03 × 4 = 0.12.
19
20. • 3. Color Keyword Search
• The system converts each color keyword into
an RGB color c* for that word using a hard-
coded lookup table. / Six-digit hexadecimal
• The current prototype supports twelve
common color words.
• To determine how well a page matches the
query color.
• It computes a score s based on the page’s
histogram h. h is a set of pairs (ci; fi),
• ci - RGB color and fi - fraction of pixels in the
page whose nearest histogram color is ci.
S=S1+S2+S3……..+Sh
20
21. • frequently-occurring colors that are close to
query color should have a high score
contribution si.
• use a value of one-third the diagonal
length of the RGB color cube for the threshold.
• The system returns as search results only the
pages with s greater than a heuristically
determined threshold.
fraction of pixels
RGB color
each histogram pair (ci; fi)
21
23. 4. Style
Keyword
Search
The current
prototype
invokes these
rules when the
user searches for
the bold terms.
35 terms—
including
synonyms—map
to 14 unique
rules.
All the features
in any single rule
are weighted
equally. 23
24. • EVALUATION
• Method
• An online study instructed participants to find
a set of inspirational design examples.
• Participants were randomly assigned to use
either d.tour or a Web search engine to find
examples.
• Provide feedback about their search strategies
and experience.
24
25. • Participants
• 40 participants: 20 through student email lists
at university and 20 more via Amazon
Mechanical Turk.
• Participants searched for design examples.
• Finishing search, responded to a
questionnaire.
• Provided four inspirational design examples
URLs and explanations for their chosen
examples.
25
26. • Results
• Mainly reported querying with keywords such
as ‘science,’ ‘kids,’ and ‘education.’
• Retrieved pages from their own mental search
history, recalling well-designed, potentially
relevant sites they had seen recently.
• Lack of style-based search or the insufficiency
of searching by page content as serious
roadblocks.
• Search function “rated sites by mood or feel.”
• “Google is only really useful for searching for
actual page content, not page style.
26
27. • d.tour as a “really good tool
to find quick examples.”
• Specific interface features as
useful.
• Current Query as reminder
of search history.
• ‘Ambient Style Display’
received mixed feedback,
difficulty interpreting and
operationalizing the
information it provided.
• Low corpus size as a
limitation for d.tour.
27
28. • Participants
using d.tour
focused on
the style
and
present-
ction of
their
chosen
designs.
28
29. • Many
participants
using
search
engines
chose
pages on a
topic
relevant to
the design
task.
29
30. • CONCLUSION AND FUTURE WORK
• 1. d.tour currently standardizes all page
features, effectively giving them all the same
weight.
• 2. Expanding the stylistic vocabulary. Apply
machine learning techniques to a large corpus of
stylistically-tagged Web pages.
• 3. Many of d.tour’s features rely on the HTML
DOM, which simplifies a page to a hierarchy of
rectangles.
• 4. d.tour does not attempt to capture non-static
aspects of design such as animation or
interaction.
• 5. A full-fledged system should scale to much
large page databases—perhaps even the entire 30
31. 3. The wine bottles and the type on this page overlap and span
multiple DOM nodes (outlined in blue). Breaking the strong grid lines
specified by the DOM. d.tour’s current feature set cannot support
31
queries about ‘organic’ page layouts such as this.